extern "C" OFFLOAD OFFLOAD_TARGET_ACQUIRE(
TARGET_TYPE target_type,
int target_number,
int is_optional,
_Offload_status* status,
const char* file,
uint64_t line
)
{
bool retval;
OFFLOAD ofld;
// initialize status
if (status != 0) {
status->result = OFFLOAD_UNAVAILABLE;
status->device_number = -1;
status->data_sent = 0;
status->data_received = 0;
}
// make sure libray is initialized
retval = __offload_init_library();
// OFFLOAD_TIMER_INIT must follow call to __offload_init_library
OffloadHostTimerData * timer_data = OFFLOAD_TIMER_INIT(file, line);
OFFLOAD_TIMER_START(timer_data, c_offload_host_total_offload);
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
// initialize all devices is init_type is on_offload_all
if (retval && __offload_init_type == c_init_on_offload_all) {
for (int i = 0; i < mic_engines_total; i++) {
mic_engines[i].init();
}
}
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
OFFLOAD_TIMER_START(timer_data, c_offload_host_target_acquire);
if (target_type == TARGET_HOST) {
// Host always available
retval = true;
}
else if (target_type == TARGET_MIC) {
if (target_number >= -1) {
if (retval) {
if (target_number >= 0) {
// User provided the device number
target_number = target_number % mic_engines_total;
}
else {
// use device 0
target_number = 0;
}
// reserve device in ORSL
if (is_optional) {
if (!ORSL::try_reserve(target_number)) {
target_number = -1;
}
}
else {
if (!ORSL::reserve(target_number)) {
target_number = -1;
}
}
// initialize device
if (target_number >= 0 &&
__offload_init_type == c_init_on_offload) {
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
mic_engines[target_number].init();
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
}
}
else {
// fallback to CPU
target_number = -1;
}
if (target_number < 0 || !retval) {
if (!is_optional && status == 0) {
LIBOFFLOAD_ERROR(c_device_is_not_available);
exit(1);
}
retval = false;
}
}
else {
LIBOFFLOAD_ERROR(c_invalid_device_number);
exit(1);
}
}
if (retval) {
ofld = new OffloadDescriptor(target_number, status,
!is_optional, false, timer_data);
OFFLOAD_TIMER_HOST_MIC_NUM(timer_data, target_number);
Offload_Report_Prolog(timer_data);
OFFLOAD_DEBUG_TRACE_1(2, timer_data->offload_number, c_offload_start,
"Starting offload: target_type = %d, "
"number = %d, is_optional = %d\n",
target_type, target_number, is_optional);
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire);
}
else {
ofld = NULL;
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire);
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_total_offload);
offload_report_free_data(timer_data);
}
return ofld;
}
void OffloadDescriptor::offload(
uint32_t buffer_count,
void** buffers,
void* misc_data,
uint16_t misc_data_len,
void* return_data,
uint16_t return_data_len
)
{
FunctionDescriptor *func = (FunctionDescriptor*) misc_data;
const char *name = func->data;
OffloadDescriptor ofld;
char *in_data = 0;
char *out_data = 0;
char *timer_data = 0;
console_enabled = func->console_enabled;
timer_enabled = func->timer_enabled;
offload_report_level = func->offload_report_level;
offload_number = func->offload_number;
ofld.set_offload_number(func->offload_number);
#ifdef SEP_SUPPORT
if (sep_monitor) {
if (__sync_fetch_and_add(&sep_counter, 1) == 0) {
OFFLOAD_DEBUG_TRACE(2, "VTResumeSampling\n");
VTResumeSampling();
}
}
#endif // SEP_SUPPORT
OFFLOAD_DEBUG_TRACE_1(2, ofld.get_offload_number(),
c_offload_start_target_func,
"Offload \"%s\" started\n", name);
// initialize timer data
OFFLOAD_TIMER_INIT();
OFFLOAD_TIMER_START(c_offload_target_total_time);
OFFLOAD_TIMER_START(c_offload_target_descriptor_setup);
// get input/output buffer addresses
if (func->in_datalen > 0 || func->out_datalen > 0) {
if (func->data_offset != 0) {
in_data = (char*) misc_data + func->data_offset;
out_data = (char*) return_data;
}
else {
char *inout_buf = (char*) buffers[--buffer_count];
in_data = inout_buf;
out_data = inout_buf;
}
}
// assign variable descriptors
ofld.m_vars_total = func->vars_num;
if (ofld.m_vars_total > 0) {
uint64_t var_data_len = ofld.m_vars_total * sizeof(VarDesc);
ofld.m_vars = (VarDesc*) malloc(var_data_len);
if (ofld.m_vars == NULL)
LIBOFFLOAD_ERROR(c_malloc);
memcpy(ofld.m_vars, in_data, var_data_len);
in_data += var_data_len;
func->in_datalen -= var_data_len;
}
// timer data
if (func->timer_enabled) {
uint64_t timer_data_len = OFFLOAD_TIMER_DATALEN();
timer_data = out_data;
out_data += timer_data_len;
func->out_datalen -= timer_data_len;
}
// init Marshallers
ofld.m_in.init_buffer(in_data, func->in_datalen);
ofld.m_out.init_buffer(out_data, func->out_datalen);
// copy buffers to offload descriptor
std::copy(buffers, buffers + buffer_count,
std::back_inserter(ofld.m_buffers));
OFFLOAD_TIMER_STOP(c_offload_target_descriptor_setup);
// find offload entry address
OFFLOAD_TIMER_START(c_offload_target_func_lookup);
offload_func_with_parms entry = (offload_func_with_parms)
__offload_entries.find_addr(name);
if (entry == NULL) {
#if OFFLOAD_DEBUG > 0
if (console_enabled > 2) {
__offload_entries.dump();
}
#endif
LIBOFFLOAD_ERROR(c_offload_descriptor_offload, name);
exit(1);
}
OFFLOAD_TIMER_STOP(c_offload_target_func_lookup);
OFFLOAD_TIMER_START(c_offload_target_func_time);
// execute offload entry
entry(&ofld);
OFFLOAD_TIMER_STOP(c_offload_target_func_time);
OFFLOAD_TIMER_STOP(c_offload_target_total_time);
// copy timer data to the buffer
OFFLOAD_TIMER_TARGET_DATA(timer_data);
OFFLOAD_DEBUG_TRACE(2, "Offload \"%s\" finished\n", name);
#ifdef SEP_SUPPORT
if (sep_monitor) {
if (__sync_sub_and_fetch(&sep_counter, 1) == 0) {
OFFLOAD_DEBUG_TRACE(2, "VTPauseSampling\n");
VTPauseSampling();
}
}
#endif // SEP_SUPPORT
}
extern "C" OFFLOAD OFFLOAD_TARGET_ACQUIRE1(
const int* device_num,
const char* file,
uint64_t line
)
{
int target_number;
// make sure libray is initialized and at least one device is available
if (!__offload_init_library()) {
LIBOFFLOAD_ERROR(c_device_is_not_available);
exit(1);
}
// OFFLOAD_TIMER_INIT must follow call to __offload_init_library
OffloadHostTimerData * timer_data = OFFLOAD_TIMER_INIT(file, line);
OFFLOAD_TIMER_START(timer_data, c_offload_host_total_offload);
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
if (__offload_init_type == c_init_on_offload_all) {
for (int i = 0; i < mic_engines_total; i++) {
mic_engines[i].init();
}
}
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
OFFLOAD_TIMER_START(timer_data, c_offload_host_target_acquire);
// use default device number if it is not provided
if (device_num != 0) {
target_number = *device_num;
}
else {
target_number = __omp_device_num;
}
// device number should be a non-negative integer value
if (target_number < 0) {
LIBOFFLOAD_ERROR(c_omp_invalid_device_num);
exit(1);
}
// should we do this for OpenMP?
target_number %= mic_engines_total;
// reserve device in ORSL
if (!ORSL::reserve(target_number)) {
LIBOFFLOAD_ERROR(c_device_is_not_available);
exit(1);
}
// initialize device(s)
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
if (__offload_init_type == c_init_on_offload) {
mic_engines[target_number].init();
}
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
OFFLOAD ofld =
new OffloadDescriptor(target_number, 0, true, true, timer_data);
OFFLOAD_TIMER_HOST_MIC_NUM(timer_data, target_number);
Offload_Report_Prolog(timer_data);
OFFLOAD_DEBUG_TRACE_1(2, timer_data->offload_number, c_offload_start,
"Starting OpenMP offload, device = %d\n",
target_number);
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire);
return ofld;
}
